app.py

import gradio as gr
from functools import lru_cache
from hffs.fs import HfFileSystem
from typing import List, Tuple, Callable
import pandas as pd
import numpy as np
import pyarrow as pa
import pyarrow.parquet as pq
from functools import partial
from tqdm.contrib.concurrent import thread_map
from datasets import Features, Image, Audio
from fastapi import FastAPI, Response
import uvicorn
import os


class AppError(RuntimeError):
    pass


APP_URL = "http://127.0.0.1:7860" if os.getenv("DEV") else "https://lhoestq-datasets-explorer.hf.space"
PAGE_SIZE = 20
MAX_CACHED_BLOBS = PAGE_SIZE * 10
_blobs_cache = {}

#####################################################
# Define routes for image and audio files
#####################################################

app = FastAPI()


@app.get(
    "/image",
    responses={200: {"content": {"image/png": {}}}},
    response_class=Response,
)
def image(id: str):
    blob = get_blob(id)
    return Response(content=blob, media_type="image/png")


@app.get(
    "/audio",
    responses={200: {"content": {"audio/wav": {}}}},
    response_class=Response,
)
def audio(id: str):
    blob = get_blob(id)
    return Response(content=blob, media_type="audio/wav")


def push_blob(blob: bytes, blob_id: str) -> str:
    global _blobs_cache
    if blob_id in _blobs_cache:
        del _blobs_cache[blob_id]
    _blobs_cache[blob_id] = blob
    if len(_blobs_cache) > MAX_CACHED_BLOBS:
        del _blobs_cache[next(iter(_blobs_cache))]
    return blob_id


def get_blob(blob_id: str) -> bytes:
    global _blobs_cache
    return _blobs_cache[blob_id]


def blobs_to_urls(blobs: List[bytes], type: str, prefix: str) -> List[str]:
    image_blob_ids = [push_blob(blob, f"{prefix}-{i}") for i, blob in enumerate(blobs)]
    return [APP_URL + f"/{type}?id={blob_id}" for blob_id in image_blob_ids]


#####################################################
# List configs, splits and parquet files
#####################################################


@lru_cache(maxsize=128)
def get_parquet_fs(dataset: str) -> HfFileSystem:
    try:
        fs = HfFileSystem(dataset, repo_type="dataset", revision="refs/convert/parquet")
        if any(fs.isfile(path) for path in fs.ls("") if not path.startswith(".")):
            raise AppError(f"Parquet export doesn't exist for '{dataset}'.")
        return fs
    except:
        raise AppError(f"Parquet export doesn't exist for '{dataset}'.")


@lru_cache(maxsize=128)
def get_parquet_configs(dataset: str) -> List[str]:
    fs = get_parquet_fs(dataset)
    return [path for path in fs.ls("") if fs.isdir(path)]


def _sorted_split_key(split: str) -> str:
    return split if not split.startswith("train") else chr(0) + split  # always "train" first


@lru_cache(maxsize=128)
def get_parquet_splits(dataset: str, config: str) -> List[str]:
    fs = get_parquet_fs(dataset)
    all_parts = [path.rsplit(".", 1)[0].split("-") for path in fs.glob(f"{config}/*.parquet")]
    return sorted(set(parts[-4] if len(parts) > 3 and parts[-2] == "of" else parts[-1] for parts in all_parts), key=_sorted_split_key)


#####################################################
# Index and query Parquet data
#####################################################


RowGroupReaders = List[Callable[[], pa.Table]]


@lru_cache(maxsize=128)
def index(dataset: str, config: str, split: str) -> Tuple[np.ndarray, RowGroupReaders, int, Features]:
    fs = get_parquet_fs(dataset)
    sources = fs.glob(f"{config}/*-{split}.parquet") + fs.glob(f"{config}/*-{split}-*-of-*.parquet")
    if not sources:
        if config not in get_parquet_configs(dataset):
            raise AppError(f"Invalid config {config}. Available configs are: {', '.join(get_parquet_configs(dataset))}.")
        else:
            raise AppError(f"Invalid split {split}. Available splits are: {', '.join(get_parquet_splits(dataset, config))}.")
    desc = f"{dataset}/{config}/{split}"
    all_pf: List[pq.ParquetFile] = thread_map(partial(pq.ParquetFile, filesystem=fs), sources, desc=desc, unit="pq")
    features = Features.from_arrow_schema(all_pf[0].schema.to_arrow_schema())
    rg_offsets = np.cumsum([pf.metadata.row_group(i).num_rows for pf in all_pf for i in range(pf.metadata.num_row_groups)])
    rg_readers = [partial(pf.read_row_group, i) for pf in all_pf for i in range(pf.metadata.num_row_groups)]
    max_page = 1 + (rg_offsets[-1] - 1) // PAGE_SIZE
    return rg_offsets, rg_readers, max_page, features


def query(page: int, page_size: int, rg_offsets: np.ndarray, rg_readers: RowGroupReaders) -> pd.DataFrame:
    start_row, end_row = (page - 1) * page_size, min(page * page_size, rg_offsets[-1] - 1) # both included
    # rg_offsets[start_rg - 1] <= start_row < rg_offsets[start_rg]
    # rg_offsets[end_rg - 1]   <=  end_row  < rg_offsets[end_rg]
    start_rg, end_rg = np.searchsorted(rg_offsets, [start_row, end_row], side="right") # both included
    pa_table = pa.concat_tables([rg_readers[i]() for i in range(start_rg, end_rg + 1)])
    offset = start_row - (rg_offsets[start_rg - 1] if start_rg > 0 else 0)
    pa_table = pa_table.slice(offset, page_size)
    return pa_table.to_pandas()


def sanitize_inputs(dataset: str, config: str, split: str, page:  str) -> Tuple[str, str, str, int]:
    try:
        page = int(page)
        assert page > 0
    except:
        raise AppError(f"Bad page: {page}")
    if not dataset:
        raise AppError("Empty dataset name")
    if not config:
        raise AppError(f"Empty config. Available configs are: {', '.join(get_parquet_configs(dataset))}.")
    if not split:
        raise AppError(f"Empty split. Available splits are: {', '.join(get_parquet_splits(dataset, config))}.")
    return dataset, config, split, int(page)


@lru_cache(maxsize=128)
def get_page_df(dataset: str, config: str, split: str, page: str) -> Tuple[pd.DataFrame, int, Features]:
    dataset, config, split, page = sanitize_inputs(dataset, config, split, page)
    rg_offsets, rg_readers, max_page, features = index(dataset, config, split)
    if page > max_page:
        raise AppError(f"Page {page} does not exist")
    df = query(page, PAGE_SIZE, rg_offsets=rg_offsets, rg_readers=rg_readers)
    return df, max_page, features


#####################################################
# Format results
#####################################################


def get_page(dataset: str, config: str, split: str, page: str) -> Tuple[str, int, str]:
    df, max_page, features = get_page_df(dataset, config, split, page)
    unsupported_columns = []
    for column, feature in features.items():
        if isinstance(feature, Image):
            blob_type = "image"  # TODO: support audio - right now it seems that the markdown renderer in gradio doesn't support audio and shows nothing
            blob_urls = blobs_to_urls([item.get("bytes") if isinstance(item, dict) else None for item in df[column]], blob_type, prefix=f"{dataset}-{config}-{split}-{page}-{column}")
            df = df.drop([column], axis=1)
            df[column] = [f"![]({url})" for url in blob_urls]
        elif any(bad_type in str(feature) for bad_type in ["Image(", "Audio(", "'binary'"]):
            unsupported_columns.append(column)
            df = df.drop([column], axis=1)
    info = "" if not unsupported_columns else f"Some columns are not supported yet: {unsupported_columns}"
    return df.to_markdown(index=False), max_page, info


#####################################################
# Gradio app
#####################################################


with gr.Blocks() as demo:
    gr.Markdown("# 📖 Datasets Explorer\n\nAccess any slice of data of any dataset on the [Hugging Face Dataset Hub](https://huggingface.co/datasets)")
    cp_dataset = gr.Textbox("frgfm/imagenette", label="Pick a dataset", placeholder="competitions/aiornot")
    cp_go = gr.Button("Explore")
    cp_config = gr.Dropdown(["plain_text"], value="plain_text", label="Config", visible=False)
    cp_split = gr.Dropdown(["train", "validation"], value="train", label="Split", visible=False)
    cp_goto_next_page = gr.Button("Next page", visible=False)
    cp_error = gr.Markdown("", visible=False)
    cp_info = gr.Markdown("", visible=False)
    cp_result = gr.Markdown("", visible=False)
    with gr.Row():
        cp_page = gr.Textbox("1", label="Page", placeholder="1", visible=False)
        cp_goto_page = gr.Button("Go to page", visible=False)

    def show_error(message: str) -> dict():
        return {
            cp_error: gr.update(visible=True, value=f"## ❌ Error:\n\n{message}"),
            cp_info: gr.update(visible=False, value=""),
            cp_result: gr.update(visible=False, value=""),
        }

    def show_dataset_at_config_and_split_and_page(dataset: str, config: str, split: str, page: str) -> dict:
        try:
            markdown_result, max_page, info = get_page(dataset, config, split, page)
            info = f"({info})" if info else ""
            return {
                cp_result: gr.update(visible=True, value=markdown_result),
                cp_info: gr.update(visible=True, value=f"Page {page}/{max_page} {info}"),
                cp_error: gr.update(visible=False, value="")
            }
        except AppError as err:
            return show_error(str(err))
    
    def show_dataset_at_config_and_split_and_next_page(dataset: str, config: str, split: str, page: str) -> dict:
        try:
            next_page = str(int(page) + 1)
            return {
                **show_dataset_at_config_and_split_and_page(dataset, config, split, next_page),
                cp_page: gr.update(value=next_page, visible=True),
            }
        except AppError as err:
            return show_error(str(err))

    def show_dataset_at_config_and_split(dataset: str, config: str, split: str) -> dict:
        try:
            return {
                **show_dataset_at_config_and_split_and_page(dataset, config, split, "1"),
                cp_page: gr.update(value="1", visible=True),
                cp_goto_page: gr.update(visible=True),
                cp_goto_next_page: gr.update(visible=True),
            }
        except AppError as err:
            return show_error(str(err))

    def show_dataset_at_config(dataset: str, config: str) -> dict:
        try:
            splits = get_parquet_splits(dataset, config)
            if not splits:
                raise AppError(f"Dataset {dataset} with config {config} has no splits.")
            else:
                split = splits[0]
            return {
                **show_dataset_at_config_and_split(dataset, config, split),
                cp_split: gr.update(value=split, choices=splits, visible=len(splits) > 1),
            }
        except AppError as err:
            return show_error(str(err))

    def show_dataset(dataset: str) -> dict:
        try:
            configs = get_parquet_configs(dataset)
            if not configs:
                raise AppError(f"Dataset {dataset} has no configs.")
            else:
                config = configs[0]
            return {
                **show_dataset_at_config(dataset, config),
                cp_config: gr.update(value=config, choices=configs, visible=len(configs) > 1),
            }
        except AppError as err:
            return show_error(str(err))

    all_outputs = [cp_config, cp_split, cp_page, cp_goto_page, cp_goto_next_page, cp_result, cp_info, cp_error]
    cp_go.click(show_dataset, inputs=[cp_dataset], outputs=all_outputs)
    cp_config.change(show_dataset_at_config, inputs=[cp_dataset, cp_config], outputs=all_outputs)
    cp_split.change(show_dataset_at_config_and_split, inputs=[cp_dataset, cp_config, cp_split], outputs=all_outputs)
    cp_goto_page.click(show_dataset_at_config_and_split_and_page, inputs=[cp_dataset, cp_config, cp_split, cp_page], outputs=all_outputs)
    cp_goto_next_page.click(show_dataset_at_config_and_split_and_next_page, inputs=[cp_dataset, cp_config, cp_split, cp_page], outputs=all_outputs)


if __name__ == "__main__":

    app = gr.mount_gradio_app(app, demo, path="/", gradio_api_url="http://localhost:7861/")
    uvicorn.run(app, host="0.0.0.0", port=7860)